Use ChatGPT to Argue Against Your Own Decisions

Two years ago I made a hiring decision I was completely confident about. The candidate checked every box. My gut said yes. My team said yes. Three months later I was undoing the damage. The decision was bad in ways I could have spotted beforehand if I had been willing to look. That experience sent me looking for a forcing function that would make me actually look. ChatGPT became that forcing function, and the technique I stumbled onto has since changed how I approach every decision that matters.

The idea is simple: before I commit to anything significant, I ask ChatGPT to argue against it as hard as it can. Not to validate me. Not to help me feel better. To find every credible reason why I might be wrong.

Why Your Brain Is Working Against You

Human decision-making is riddled with systematic errors. Confirmation bias makes you seek out information that supports what you already believe. Optimism bias makes you underestimate how often things go wrong. Sunk cost fallacy keeps you committed to paths that stopped making sense. The planning fallacy makes every timeline feel achievable until it isn’t.

These aren’t flaws you can think your way out of. They’re features of how cognition works under uncertainty. What you need is something external that doesn’t share your emotional investment in the outcome, doesn’t care about your ego, and has no incentive to tell you what you want to hear.

ChatGPT, used correctly, is exactly that.

The key phrase is “used correctly.” Asking ChatGPT “is this a good idea?” will get you a balanced, hedged, diplomatic non-answer. That’s useless. You need a structured prompt that forces it into adversarial mode, and a conversation framework that extracts the analysis worth having.

The Five-Step Framework

This is the workflow I run for any decision with meaningful consequences: a hire, a product launch, a strategic pivot, a large purchase, a partnership. It takes 10 to 20 minutes and has saved me from at least four decisions I would have regretted.

Step 1: State Your Decision Clearly

Start a fresh conversation and lay out the decision in full detail. Don’t editorialize. Don’t explain why it’s a good idea. Just describe what you’re planning to do, the context, the constraints, and the alternatives you’ve already ruled out.

The more honest and complete this briefing is, the more useful the adversarial analysis will be. If you omit inconvenient facts, you’re gaming the system against yourself.

Prompt template:

I'm about to make the following decision: [describe decision in detail].
Context: [relevant background — resources, constraints, timeline, alternatives considered].
My reasoning is: [why you currently think this is the right call].
Do not evaluate this yet. Just confirm you understand it.

Getting a confirmation before proceeding forces ChatGPT to demonstrate comprehension. If it’s missed something important, you catch it here.

Step 2: The Devil’s Advocate Request

This is the core of the technique. You’re asking ChatGPT to steelman the case against your decision. Not a half-hearted list of risks, but the most credible, rigorous argument that you are wrong.

Prompt template:

Now I want you to argue against this decision as forcefully as possible. 
Steelman the opposing view. Assume a smart, experienced person who has seen similar situations 
go wrong would look at my plan and immediately spot the problems I'm not seeing.
What would they say? Be specific. Do not soften the critique. Do not balance it with positives.
I want the strongest possible case that I should not do this.

The instruction to “not soften the critique” and “not balance it with positives” is critical. By default, ChatGPT will hedge. This prompt overrides that instinct.

Step 3: Bias Identification

After the devil’s advocate critique, ask explicitly about cognitive biases. This step often surfaces the most uncomfortable and most valuable insights.

Prompt template:

Based on what I've told you, which cognitive biases are most likely distorting my judgment here?
Be specific about how each bias is showing up in my stated reasoning.
Common ones to consider: confirmation bias, optimism bias, sunk cost fallacy, planning fallacy, 
status quo bias, overconfidence effect. But don't limit yourself to this list.

When I ran this on the hiring decision I mentioned earlier, the response identified optimism bias (I was projecting best-case performance) and in-group favoritism (the candidate reminded me of myself at an earlier career stage). Both were accurate. Neither would I have named unprompted.

Step 4: The Pre-Mortem

Gary Klein developed the pre-mortem technique at the Army Research Institute. The idea: before a project starts, imagine it has already failed spectacularly, then work backward to explain why. It forces your brain out of “how do I make this work” mode and into “what would actually go wrong” mode.

ChatGPT runs an excellent pre-mortem.

Prompt template:

Run a pre-mortem on this decision. Assume it's 18 months from now and this decision turned out 
to be a serious mistake. Write a short post-mortem report: what went wrong, why, 
and what warning signs were present at the time of the decision that were ignored or minimized.

The 18-month framing matters. It’s close enough to feel real, far enough to expose compounding problems that don’t surface immediately.

Step 5: The Steelman in Favor (Then Your Decision)

Only after you’ve done steps 2 through 4 do you ask ChatGPT to make the affirmative case. This order is deliberate. If you start with the positive case, your confirmation bias kicks in and you stop listening to the critique.

Prompt template:

Now steelman the case FOR this decision. Same rigor as before. 
What's the strongest argument that this is actually the right call?

Read both steelmans back to back. Then make your decision. You will have more information than you started with, and you’ll have been forced to confront arguments you would otherwise have talked yourself past.

What Good Output Actually Looks Like

Here’s a condensed example. I was considering shutting down a small content project that had been running for four months with below-expectations traffic numbers. My instinct was to kill it.

Devil’s advocate output (abridged):

“You may be killing this at exactly the wrong moment. SEO-driven content projects typically show meaningful traction only after six to nine months of consistent publishing, not four. Your data shows rising time-on-page and a small but growing returning visitor rate — both indicators of an audience forming, not one failing to form. Shutting down now would be a classic sunk cost reversal: the sunk cost isn’t a reason to continue, but the forward-looking metrics are. The strongest case against killing this is that you’re making an irreversible decision using four months of data on a twelve-month timeline.”

Bias identification output (abridged):

“The most likely distortion here is the planning fallacy — you set expectations for a four-month traffic ramp that most content projects can’t deliver, then concluded the project was failing when it failed to meet an unrealistic benchmark. There’s also possible availability bias: recent disappointing traffic numbers are mentally available and emotionally weighted more heavily than the slower-building positive signals.”

I didn’t shut it down. By month seven, organic traffic had tripled. The devil’s advocate was right. I was wrong.

When to Use This (And When Not To)

This technique is highest value for decisions that are:

• High stakes: the downside of being wrong is significant
• Relatively irreversible: you can’t easily course-correct once committed
• Emotionally loaded: you have a strong preference going in
• Time-pressured: there’s external pressure pushing you toward a faster decision than the situation warrants

It’s lower value for reversible, low-stakes decisions where speed matters more than rigor. Don’t run a 15-minute red-team session on which project management tool to try next week.

Which Model to Use

For this workflow, reasoning quality matters more than response speed. The base ChatGPT-4o model handles it well for most decisions. For genuinely complex, high-stakes decisions (major hires, large capital commitments, strategic pivots), use ChatGPT Plus with the o3 or o1 model. The extended reasoning chain produces significantly more rigorous adversarial analysis.

If you’re already using Claude, the same framework works with Claude Pro. Claude’s longer context window and tendency toward structured output can be advantageous for decisions with a lot of background material. For a head-to-head comparison of both platforms, see our Claude Pro vs ChatGPT Plus comparison based on four months of real use.

The underlying prompt engineering principles here are the same ones covered in our prompt engineering guide for Claude and GPT-4o. Role assignment, constraint specification, and output formatting all apply.

The Discipline Problem

The hardest part of this technique is not the prompt. It’s what you do with the output.

When ChatGPT makes a strong case against something you want to do, the natural response is to look for reasons why the critique doesn’t apply to your specific situation. This is confirmation bias reasserting itself. You’ve replaced one source of confirmation with another.

The discipline required: treat the adversarial output as a genuine challenge that requires a genuine rebuttal, not a dismissal. If you can’t articulate a specific reason why a critique doesn’t hold, it probably does.

One way to enforce this: after reading the devil’s advocate output, write a short paragraph responding to each major point. If you’re skipping points, notice that. If your response is “that’s true but I’m going ahead anyway,” at least you’re making a conscious choice rather than a blind one.

This connects to something worth reading in depth if this topic interests you. Daniel Kahneman’s Thinking, Fast and Slow remains the most accessible rigorous treatment of the cognitive biases this technique is designed to surface. It will make you better at interpreting what ChatGPT finds.

Building It Into Your Workflow

The technique works best as a habit, not an occasional experiment. Here’s how to systematize it:

1. Define your decision threshold. Decide in advance which class of decisions triggers the protocol. Mine: anything that would take more than a week to reverse if wrong.

2. Keep a decisions log. Record the decision, the key devil’s advocate arguments, which ones you dismissed and why, and the eventual outcome. Over time you’ll see which critiques you were right to dismiss and which ones you should have weighted more heavily.

3. Revisit your dismissed arguments. When a decision goes wrong, go back to the ChatGPT session. The warning was almost certainly there. This is how you calibrate.

For more on building systematic AI-assisted workflows, the 100 tips for building a personal AI agent covers a lot of the surrounding infrastructure: memory, context management, and how to get consistent outputs across sessions.

Why This Works at a Structural Level

The reason this technique has more value than simply thinking harder about a decision comes down to a structural asymmetry. When you reason about your own decision, you’re using the same cognitive apparatus that produced the decision. The biases that shaped your conclusion are the same biases you’re using to evaluate your conclusion.

ChatGPT doesn’t share those biases because it doesn’t share your history, your emotional investment, or your desire to be right. What it does share is enough general knowledge about how decisions like yours tend to fail that it can construct a credible adversarial case.

It’s not a replacement for an expert advisor who has deep domain knowledge and a track record of similar decisions. But it’s available immediately, has no social cost, won’t soften the critique to protect the relationship, and can run the analysis in five minutes at two in the morning before a deadline.

That combination of availability and zero social friction is why it’s become the most reliable tool in my decision-making process.

Affiliate disclosure: Some links in this post are affiliate links. If you click through and make a purchase, I may earn a commission at no extra cost to you.

Start With Your Next Decision

Pick the next meaningful decision on your list. Not a hypothetical, a real one you’re currently evaluating. Run the five-step framework. Write down the strongest critique that comes back and ask yourself whether you have a specific, evidence-based rebuttal or just a feeling that it doesn’t apply to you.

That moment of honest assessment is the whole technique. Everything else is just structure that gets you there.

If you’re new to prompt engineering and want to get more out of this workflow, start with the prompt engineering guide to sharpen how you frame the briefing in step one. Better input, better adversarial output.